DeepSeek Harness Adds Multimodal Image Input — the Harness Got Eyes, the Models Didn't

What shipped, per official sources

Background: Harness v0.1 was open-sourced Aug 13 alongside V4-Pro GA — see our v0.1 launch brief for the architecture (everything-is-a-plugin on Cordis, four work modes, MIT license).

The split: harness plumbing vs. model capability

• What got multimodal is the runtime, not the models. Paste an image in dsh today and it persists across MCP/ACP tool calls, forwards through PTC-mode programs, and can ride along in /goal and /plan planning commands. That is real plumbing that was missing at v0.1 — early reviewers dinged the launch-week stack precisely because a text-only model couldn't chase a two-line visual bug.
• DeepSeek's own API models were still text-only at publication — and gained vision a day later. On Aug 21, 2026, DeepSeek officially launched deepseek-v4-flash-vision-exp, an experimental vision model in the Flash family at unchanged pricing; V4-Pro remains text-only. At publication the changelog had carried no vision entry since the Aug 13 V4-Pro GA note, and a top Hugging Face discussion on V4-Pro complained it lacked "native multimodality every other Chinese flagship model has now, including Qwen and Kimi" — exactly the gap the launch just closed for Flash.
• The official docs make the intended pattern explicit. The repo's troubleshooting table: if an image is refused before sending, the model "declares no image modality" — fix it by adding input: [text, image] to the model entry in $DSH_HOME/settings.yaml. In other words: flip on image requests for a model that actually has vision, via DeepSeek's adapter or any OpenAI-compatible endpoint.
• The vision gap was already being bridged by plugins. modlens (self-described first vision plugin for dsh) lets you paste an image into a text-only session and get structured JSON evidence back — OCR, layout, semantics — routed through a pool of vision engines (a free Gemini key, Antigravity CLI, or any OpenAI-compatible endpoint; its README demos Qwen-VL via DashScope's compatible mode). Similar community tools include agent-vision-toolkit (image Q&A, long-screenshot OCR, UI restoration, GUI automation) and dsh-vision-router. Community projects, unaffiliated with DeepSeek — quality varies.

What the overseas coverage says

• Turing Post (Aug 17) published the sharpest English-language take: "DeepSeek is having its second DeepSeek moment." Its argument: with R1, DeepSeek weakened the model moat; with Harness (MIT, 149K+ stars at publication), it is "opening up the layer everyone had started looking to as the next moat" — the agent execution layer. It also traces Cordis's pedigree to Koishi, the chatbot framework by Shigma (now at DeepSeek), whose four years of plugin-lifecycle problems turned out to be agent-runtime problems.
• Launch-week press (covered in our v0.1 brief): VentureBeat framed dsh as an "open-source rival to Claude Code"; The Decoder profiled the architecture and project lead Cui Tianyi (ex-Jane Street); 36Kr's overnight hands-on found 288 plugin repos within 24 hours; Pandaily organized it as Model + Harness = Agent.
• The rc.7/rc.8 releases themselves were mostly tracked by changelog watchers and the plugin ecosystem rather than mainstream outlets in the first 48 hours — third-party guides (chat-deep.ai, dshdocs.com, freedom.tech) logged the "durable image attachments" line within a day. If you follow the story, follow the release notes, not the headlines.
• On X, the project's announcement thread (@deepseek_ai, Aug 13) and maintainer posts from @tianyi remain the canonical channels; modlens author @liustack ships release notes on X first — consistent with the ecosystem's center of gravity sitting in the repo and on X rather than in blogs.

Why this matters for a China-model stack

• The harness is free; image tokens are the new variable. With V4-Pro on peak/off-peak pricing (off-peak $0.66 input / $1.98 output per 1M, peak $1.32 / $3.96 — per The Decoder's reading of the official pricing page), the cost of an agent loop shifts to wherever images get processed. Routing vision sub-calls to a cheap vision-capable model while text models drive the coding loop is exactly the multi-model pattern dsh's plugin model layer was built for.
• rc.8's gateway fix is directly relevant to relay users. The release fixes custom OpenAI-compatible gateways that failed on request-format differences or dropped reasoning content — the class of bug that bites anyone pointing dsh at a third-party endpoint rather than first-party keys. Test your endpoint against rc.8 before blaming the harness.
• The breaking-change cadence is real and delivered. v0.1 promised compatibility-breaking changes; rc.8 delivers one (SQLite storage format incompatible). Pin your dsh version, keep session data out of the upgrade path, and re-verify after every rc bump — the line is still racing toward a stable 0.1.0.
• Branding is now formal. rc.8's notes clarify that "DeepSeek Harness" is a registered trademark with published brand guidelines — plugin authors and tool builders should read them before naming derivatives.
• Independent-project framing. dsh is DeepSeek's open-source project. ChinaModelAPI is an independent OpenAI-compatible relay with no official relationship to it or to DeepSeek — the practical overlap is that a model-agnostic harness is a natural client for multi-model routing, and vision-capable model IDs should be verified live in your provider's dashboard before you wire them in.

Primary sources

• GitHub — deepseek-ai/deepseek-harness releases (rc.7 and rc.8 notes, CN/EN)
• npm — @deepseek-ai/dsh dist-tags (latest = rc.7, next = rc.8, checked 2026-08-20)
• deepseek.com/harness — official developer preview page
• DeepSeek API changelog — no vision entry since 2026-08-13 (text-only status of V4 models)
• Turing Post — FOD#163: DeepSeek is having its second DeepSeek moment (2026-08-17)
• VentureBeat — DeepSeek Harness launches as open-source rival to Claude Code (launch week)
• The Decoder — improved V4 Pro, higher API prices, open-sourced agent software (pricing detail)
• GitHub — liustack/modlens (community vision plugin for dsh, MIT)
• DeepSeek Code directory — agent-vision-toolkit plugin profile (community)
• Hugging Face discussion — V4-Pro native multimodality gap (community)

FAQ